Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for disaggregating latent causes for computer system optimization. In one aspect, a method includes accessing a data stream for data values resulting from operations performed by a computer system; providing the data values as input to a data disaggregation machine learning model that generates descriptors of latent causes of the data values; providing the data values and the descriptors of the latent causes of the data values as inputs to a control system model that generates embedded representations of commands to modify the operations performed by the computer system; determining commands to modify the operations performed by the computer system based on the embedded representations of commands to modify the operations performed by the computer system; and providing the commands to the computer system.
Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A method performed by one or more data processing apparatus, the method comprising, at each of a plurality of time steps: accessing a data stream to obtain one or more data values corresponding to the time step that result from operations performed by a computer system of one or more computers, the data values being descriptive of the operations performed by the computer system; processing an input comprising the data values corresponding to the time step using a data disaggregation machine learning model and in accordance with a set of data disaggregation machine learning model parameters of the data disaggregation machine learning model to generate a compressed representation of the data values corresponding to the time step; jointly processing: (i) the data values corresponding to the time step, and (ii) the compressed representation of the data values corresponding to the time step, using a control system model and in accordance with a set of control system model parameters to generate an output specifying commands to modify the operations performed by the computer system; and providing, to the computer system, the commands to modify the operations performed by the computer system.
This invention relates to real-time monitoring and control of computer systems using machine learning. The problem addressed is efficiently processing high-volume data streams from computer systems to optimize performance and resource usage. The method involves analyzing data streams from a computer system at multiple time steps. At each time step, data values descriptive of the system's operations are obtained. These values are processed by a data disaggregation machine learning model, which compresses the data into a compact representation while preserving relevant information. The original data and its compressed form are then jointly processed by a control system model, which generates commands to adjust the computer system's operations. These commands are sent to the system to modify its behavior dynamically. The approach enables real-time decision-making by leveraging machine learning to handle large-scale data efficiently while maintaining control over system performance. The system models and parameters are adjusted to optimize the control process, ensuring adaptive and responsive management of the computer system's operations.
2. The method of claim 1 , further comprising training the data disaggregation machine learning model using log data that stores data values resulting from operations previously performed by the computer system.
A method for improving data disaggregation in computer systems involves training a machine learning model to analyze log data generated by the system. The log data contains records of operations previously performed by the computer system, including the data values produced by those operations. By training the model on this historical log data, the system can learn to accurately separate or disaggregate data into its constituent components, improving data analysis and processing efficiency. The method may include preprocessing the log data to extract relevant features, such as timestamps, operation types, and data value patterns, before feeding it into the machine learning model. The trained model can then be applied to new data to disaggregate it into meaningful subsets, enabling more precise data handling and decision-making. This approach is particularly useful in systems where raw data is complex or aggregated, and where accurate disaggregation is critical for performance or analysis. The method may also involve validating the model's performance using additional log data to ensure accuracy and reliability.
3. The method of claim 1 , further comprising training the data disaggregation machine learning model using data values from the data stream, and wherein accessing the data stream comprises receiving the data stream from the computer system.
This invention relates to data disaggregation using machine learning, specifically addressing the challenge of extracting meaningful, granular data from aggregated or noisy data streams in real-time. The method involves training a machine learning model to process and disaggregate data values from a continuous data stream, improving accuracy and efficiency in data analysis. The model is trained using historical or real-time data values from the same data stream, allowing it to adapt to the specific characteristics of the input data. The data stream is received directly from a computer system, ensuring seamless integration with existing data sources. The disaggregation process involves decomposing the aggregated data into its constituent components, enabling more detailed insights and analysis. This approach is particularly useful in applications where raw data is noisy or aggregated, such as sensor networks, financial transactions, or energy consumption monitoring. The trained model can dynamically adjust to changes in the data stream, maintaining high accuracy over time. The invention improves upon traditional disaggregation techniques by leveraging machine learning to handle complex, real-world data scenarios.
4. The method of claim 1 , wherein the data disaggregation machine learning model is at least partially implemented in hardware.
A system and method for data disaggregation using machine learning involves processing raw sensor data to identify and separate individual energy consumption events from aggregated sensor readings. The technology addresses the challenge of analyzing energy usage in buildings or industrial settings where multiple devices contribute to a single sensor measurement, making it difficult to isolate specific consumption patterns. The machine learning model is trained on labeled data to recognize distinct energy signatures associated with different devices or activities, enabling accurate disaggregation of the aggregated sensor data into individual components. This allows for detailed energy monitoring, cost allocation, and efficiency analysis. The model can be implemented in hardware, such as specialized processors or field-programmable gate arrays (FPGAs), to improve processing speed and efficiency. Hardware implementation may involve parallel processing, dedicated circuits for feature extraction, or optimized algorithms for real-time data analysis. The system may also include data preprocessing steps to filter noise and normalize inputs, as well as post-processing to refine disaggregated results. The hardware-accelerated model enables faster and more scalable energy disaggregation, supporting applications in smart grids, building automation, and industrial energy management.
5. The method of claim 1 , wherein the data disaggregation machine learning model comprises a variational auto-encoder.
A method for disaggregating energy consumption data using a machine learning model, specifically a variational auto-encoder (VAE). The method addresses the challenge of accurately separating individual appliance energy usage from aggregated household or building-level consumption data. Traditional approaches often struggle with noise, varying appliance behaviors, and limited training data, leading to inaccurate disaggregation. The VAE-based model improves upon these limitations by learning a probabilistic representation of energy consumption patterns, enabling more robust and precise disaggregation. The VAE encodes input data into a lower-dimensional latent space, capturing key features of appliance usage, and then decodes it to reconstruct individual appliance consumption profiles. This approach leverages unsupervised learning, reducing reliance on labeled data while maintaining high accuracy. The method is particularly useful for smart energy monitoring, demand response systems, and energy efficiency applications, where understanding individual appliance usage is critical. The VAE's ability to model uncertainty in energy consumption further enhances its reliability in real-world scenarios.
6. The method of claim 1 , wherein the data values comprise program counter values, and the operations performed by the computer system include program instruction increments.
A method for processing program counter values in a computer system involves tracking and analyzing these values to optimize program execution. The program counter values indicate the current instruction address in a program, and the computer system performs operations that include incrementing these values to advance to subsequent instructions. This method is particularly useful in debugging, profiling, or performance monitoring of software, where understanding the flow of program execution is critical. By capturing and analyzing program counter values, the system can detect anomalies, optimize instruction fetching, or improve branch prediction accuracy. The method may involve comparing program counter values against expected sequences, identifying deviations, or adjusting execution flow dynamically. This approach enhances the efficiency and reliability of program execution by ensuring correct instruction sequencing and reducing unnecessary processing overhead. The technique is applicable in various computing environments, including embedded systems, real-time processing, and high-performance computing, where precise control over program flow is essential.
7. The method of claim 1 , wherein the data values comprise memory address values, and the operations performed by the computer system include memory accesses.
This invention relates to computer systems and methods for optimizing memory access operations. The problem addressed is the inefficiency in handling memory address values during data processing, which can lead to performance bottlenecks. The invention provides a solution by performing operations on memory address values in a way that reduces latency and improves system efficiency. The method involves processing data values, where these values specifically include memory address values. The computer system performs operations on these memory address values, with a focus on memory access operations. These operations may include reading from or writing to memory locations, managing memory allocation, or optimizing memory access patterns. The system may also analyze the memory address values to predict or pre-fetch data, further enhancing performance. The invention may include additional features such as tracking memory access patterns, identifying frequently accessed memory locations, and dynamically adjusting memory access strategies based on observed usage. By optimizing how memory address values are handled, the system reduces unnecessary delays and improves overall computational efficiency. This approach is particularly useful in high-performance computing environments where memory access speed is critical.
8. The method of claim 1 , further comprising: determining a performance measure of the control system model based on the data values and the outputs specifying commands to modify the operations performed by the computer system; and adjusting a set of control system model parameters based on the performance measure.
This invention relates to optimizing control systems for computer systems. The problem addressed is improving the performance of control systems that manage operations in computer systems, such as resource allocation, task scheduling, or system behavior adjustments. The invention provides a method to evaluate and refine control system models by analyzing their outputs and adjusting model parameters accordingly. The method involves using a control system model to generate commands that modify operations in a computer system. The model processes input data values representing system states or conditions and produces outputs specifying control actions. These outputs are used to adjust the computer system's operations, such as altering resource allocation or task priorities. The method further includes determining a performance measure of the control system model by comparing the data values and the resulting outputs. This performance measure quantifies how well the model's commands improve system operations. Based on this measure, the method adjusts the control system model's parameters to enhance performance. The adjustments may involve tuning coefficients, updating model structures, or refining decision logic to better align with desired system behavior. This iterative process ensures the control system model continuously improves, leading to more efficient and effective computer system operations.
9. The method of claim 1 , wherein the control system model comprises a recurrent neural network.
A system and method for controlling a process or device using a recurrent neural network (RNN) model. The invention addresses the challenge of dynamically adapting control strategies in real-time systems where input conditions and system states change over time. Traditional control systems often rely on fixed models or simple feedback loops, which may not effectively handle complex, time-dependent behaviors. The invention improves upon prior approaches by incorporating an RNN, which is particularly suited for processing sequential data and capturing temporal dependencies. The RNN model receives input data representing the current state of the system and generates control signals to adjust the system's operation. The model is trained to optimize performance metrics such as stability, efficiency, or accuracy, depending on the application. The system may also include mechanisms for updating the RNN model based on new data, allowing it to adapt to changing conditions. This approach enables more responsive and accurate control in applications such as robotics, industrial automation, or adaptive systems where traditional control methods are insufficient. The invention may further include preprocessing steps to prepare input data for the RNN and post-processing steps to refine the generated control signals. The overall system integrates the RNN model with the physical or virtual system being controlled, forming a closed-loop control architecture.
10. The method of claim 1 , wherein the commands specified by the output of the control system model comprise commands to pre-fetch data stored at memory addresses.
A system and method for optimizing data access in computing systems, particularly for reducing latency in memory operations. The invention addresses the problem of inefficient data retrieval in computing systems, where delays in accessing required data from memory can degrade performance. The system includes a control system model that generates commands to pre-fetch data stored at specific memory addresses before the data is explicitly requested by a processor. This pre-fetching mechanism anticipates future data needs based on predictive analysis, reducing the time required to access frequently used or sequentially accessed data. The control system model may analyze patterns in data access, such as sequential reads or repeated accesses to specific memory locations, to determine optimal pre-fetching strategies. By issuing pre-fetch commands in advance, the system minimizes latency and improves overall system efficiency. The method can be applied in various computing environments, including processors, memory controllers, and storage systems, to enhance performance by reducing the time spent waiting for data retrieval. The invention ensures that critical data is available when needed, thereby optimizing computational workflows and reducing bottlenecks in data-intensive applications.
11. A system comprising: one or more computers; and one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations comprising, at each of a plurality of time steps: accessing a data stream to obtain one or more data values corresponding to the time step that result from operations performed by a computer system of one or more computers, the data values being descriptive of the operations performed by the computer system; processing an input comprising the data values corresponding to the time step using a data disaggregation machine learning model and in accordance with a set of data disaggregation machine learning model parameters of the data disaggregation machine learning model to generate a compressed representation of the data values corresponding to the time step; jointly processing: (i) the data values corresponding to the time step, and (ii) the compressed representation of the data values corresponding to the time step, using a control system model and in accordance with a set of control system model parameters to generate an output specifying commands to modify the operations performed by the computer system; and providing, to the computer system, the commands to modify the operations performed by the computer system.
The invention relates to a system for optimizing computer system operations using machine learning-based data disaggregation and control. The system addresses the challenge of efficiently managing and controlling computer system performance by processing operational data streams in real-time. At each time step, the system accesses a data stream containing values descriptive of the computer system's operations. These values are processed by a data disaggregation machine learning model, which generates a compressed representation of the data. The original data values and their compressed representation are then jointly processed by a control system model to generate commands that modify the computer system's operations. These commands are provided to the computer system to adjust its behavior dynamically. The system leverages machine learning to reduce data complexity while maintaining critical operational insights, enabling more effective control and optimization of computer system performance. The data disaggregation model and control system model operate in tandem, ensuring that the system can adapt to changing conditions and improve operational efficiency over time.
12. The system of claim 11 , wherein the operations further comprise training the data disaggregation machine learning model using log data that stores data values resulting from operations previously performed by the computer system.
This invention relates to a system for improving data disaggregation in computer systems using machine learning. The system addresses the challenge of accurately separating aggregated data into its constituent components, which is critical for tasks like resource allocation, performance monitoring, and cost analysis in computing environments. The system includes a data disaggregation machine learning model that processes input data to generate disaggregated outputs. The model is trained using log data that records data values from previous operations performed by the computer system. By leveraging historical log data, the model learns patterns and relationships in the data, enabling more precise disaggregation of new input data. The system may also include a data preprocessing module to prepare the input data for the model and a post-processing module to refine the disaggregated outputs. The training process involves feeding the log data into the model, adjusting model parameters based on the results, and iteratively improving accuracy. This approach enhances the reliability and efficiency of data disaggregation in computing systems, particularly in scenarios where detailed data is initially unavailable or obscured. The system is designed to adapt to different types of log data and computing environments, making it versatile for various applications.
13. The system of claim 11 , wherein the operations further comprise training the data disaggregation machine learning model using data values from the data stream, and wherein accessing the data stream comprises receiving the data stream from the computer system.
This invention relates to a system for processing data streams using machine learning-based disaggregation. The system addresses the challenge of extracting meaningful insights from raw, aggregated data streams by employing a trained machine learning model to decompose the data into its constituent components. The system receives a data stream from a computer system and applies a data disaggregation machine learning model to analyze the incoming data values. The model is trained using historical or real-time data values from the same data stream to improve its accuracy over time. The disaggregation process involves breaking down the aggregated data into finer-grained components, enabling more detailed analysis and decision-making. The system may also include preprocessing steps to prepare the data for disaggregation, such as normalization or noise reduction. The trained model can be updated continuously as new data is received, ensuring it adapts to changing patterns in the data stream. This approach enhances data utility by transforming raw, aggregated data into actionable insights, supporting applications in monitoring, analytics, and automation.
14. The system of claim 11 , wherein the data disaggregation machine learning model is at least partially implemented in hardware.
A system for data disaggregation uses a machine learning model to process raw data from a plurality of sources, such as sensors or devices, to generate disaggregated data outputs. The system includes a data collection module that gathers raw data from the sources and a preprocessing module that conditions the data for analysis. The machine learning model, trained to identify patterns and relationships within the raw data, processes the preprocessed data to produce disaggregated outputs, which are then transmitted to a user interface or storage system. The disaggregated data may represent individual contributions from the sources, enabling detailed analysis of energy consumption, device usage, or other metrics. The system may also include a feedback loop to refine the model based on user input or additional data. In this embodiment, the machine learning model is at least partially implemented in hardware, such as through specialized processors or application-specific integrated circuits (ASICs), to improve processing speed and efficiency. This hardware implementation may accelerate real-time data analysis and reduce computational overhead compared to purely software-based solutions. The system is particularly useful in applications requiring high-speed, accurate disaggregation of complex data streams, such as smart grid monitoring or industrial process optimization.
15. The system of claim 11 , wherein the data disaggregation machine learning model comprises a variational auto-encoder.
A system for data disaggregation uses a machine learning model to process aggregated data and estimate individual data points. The system addresses the challenge of analyzing aggregated data where individual contributions are obscured, making it difficult to extract meaningful insights. The machine learning model is trained to infer the underlying individual data points from the aggregated values, enabling more granular analysis. The data disaggregation machine learning model employs a variational auto-encoder, a type of neural network that learns to encode input data into a compressed latent representation and then decode it back to reconstruct the original data. This approach allows the model to capture complex patterns in the aggregated data and generate accurate estimates of the individual data points. The variational auto-encoder is particularly effective for this task because it can handle high-dimensional data and model the uncertainty inherent in the disaggregation process. The system may also include a data preprocessing module to prepare the aggregated data for input into the machine learning model, ensuring consistency and quality. Additionally, a post-processing module may refine the output of the model to improve accuracy or align with specific requirements. The system can be applied in various domains, such as energy consumption analysis, financial transactions, or sensor data processing, where disaggregated data provides valuable insights.
16. One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations comprising, at each of a plurality of time steps: accessing a data stream to obtain one or more data values corresponding to the time step that result from operations performed by a computer system of one or more computers, the data values being descriptive of the operations performed by the computer system; processing an input comprising the data values corresponding to the time step using a data disaggregation machine learning model and in accordance with a set of data disaggregation machine learning model parameters of the data disaggregation machine learning model to generate a compressed representation of the data values corresponding to the time step; jointly processing: (i) the data values corresponding to the time step, and (ii) the compressed representation of the data values corresponding to the time step, using to a control system model and in accordance with a set of control system model parameters to generate an output specifying commands to modify the operations performed by the computer system; and providing, to the computer system, the commands to modify the operations performed by the computer system.
This invention relates to a system for optimizing computer system operations using machine learning-based data disaggregation and control. The problem addressed is efficiently managing and controlling computer system operations by processing high-dimensional data streams in real-time to generate actionable control commands. The system processes a data stream from a computer system at multiple time steps. At each time step, it accesses data values describing the system's operations. These values are fed into a data disaggregation machine learning model, which compresses them into a lower-dimensional representation while preserving relevant information. The original data and its compressed form are then jointly processed by a control system model, which generates commands to adjust the computer system's operations. These commands are sent back to the system to modify its behavior. The data disaggregation model reduces the complexity of the input data, making it easier to analyze and control. The control system model uses both the raw and compressed data to generate optimized control signals. This approach enables real-time, adaptive management of computer systems by leveraging machine learning to handle high-dimensional operational data efficiently. The system is designed to continuously monitor and adjust operations, improving performance and resource utilization.
17. The non-transitory computer storage media of claim 16 , wherein the operations further comprise training the data disaggregation machine learning model using log data that stores data values resulting from operations previously performed by the computer system.
The invention relates to a system for training a machine learning model to disaggregate data in a computer system. The problem addressed is the difficulty of accurately separating and analyzing individual data components from aggregated log data generated by computer operations. The solution involves a non-transitory computer storage media containing instructions that, when executed, perform operations to train a data disaggregation machine learning model. The model is trained using log data that records data values from previous operations executed by the computer system. This training process enables the model to learn patterns and relationships within the aggregated data, allowing it to accurately disaggregate new data into its constituent components. The system improves data analysis by providing more granular insights from historical log data, which can be used for performance monitoring, debugging, or optimization of computer systems. The invention enhances the accuracy and efficiency of data processing by leveraging machine learning to interpret complex, aggregated data streams.
18. The non-transitory computer storage media of claim 16 , wherein the operations further comprise training the data disaggregation machine learning model using data values from the data stream, and wherein accessing the data stream comprises receiving the data stream from the computer system.
The invention relates to a system for processing and analyzing data streams using machine learning models, specifically focusing on data disaggregation. The problem addressed is the efficient and accurate separation of mixed or aggregated data into its constituent components, which is crucial for applications like energy monitoring, sensor data analysis, and other domains where raw data streams contain overlapping signals. The system involves a non-transitory computer storage medium storing instructions that, when executed, perform operations including training a data disaggregation machine learning model using data values from a data stream. The data stream is received directly from a computer system, allowing real-time or near-real-time processing. The machine learning model is trained to decompose the data stream into its individual components, improving the accuracy and usability of the disaggregated data. The operations further include accessing the data stream, which may involve receiving it from a computer system, and using the trained model to process the incoming data. The system ensures that the model is continuously updated and refined using the latest data values, enhancing its performance over time. This approach enables more precise analysis and decision-making based on the disaggregated data, addressing challenges in fields where data separation is critical.
19. The non-transitory computer storage media of claim 16 , wherein the data disaggregation machine learning model is at least partially implemented in hardware.
A system and method for data disaggregation using machine learning involves processing raw data from a plurality of sources to generate disaggregated data outputs. The system includes a data ingestion module that collects raw data from multiple sources, such as sensors, databases, or other data streams. A preprocessing module processes the raw data to remove noise, normalize values, and align temporal or spatial data points. A machine learning model, trained to recognize patterns and relationships within the raw data, then analyzes the preprocessed data to identify and separate distinct data components. The disaggregated outputs are generated by the model, which may include time-series data, categorical data, or other structured data formats. The system further includes a hardware-accelerated implementation of the machine learning model, where at least a portion of the model's computations are performed using specialized hardware, such as GPUs, FPGAs, or ASICs, to improve processing speed and efficiency. The hardware implementation may involve parallel processing, optimized data pipelines, or dedicated neural network accelerators. The system may also include a validation module to verify the accuracy of the disaggregated outputs against known benchmarks or ground truth data. The overall goal is to enhance data analysis by decomposing complex datasets into meaningful, actionable components, particularly in applications such as energy monitoring, environmental sensing, or industrial automation.
20. The non-transitory computer storage media of claim 16 , wherein the data disaggregation machine learning model comprises a variational auto-encoder.
The invention relates to data disaggregation using machine learning, specifically addressing the challenge of accurately separating aggregated data into its constituent components. The system employs a machine learning model to process aggregated data, such as energy consumption measurements, and decompose it into individual sub-metered data points. The model is trained on historical data to learn patterns and relationships between aggregated and disaggregated values, enabling it to predict the underlying components of new, unseen data. A key aspect of the invention is the use of a variational auto-encoder (VAE) as the machine learning model. The VAE is a type of neural network that encodes input data into a lower-dimensional latent space and then decodes it back into the original or similar data, effectively learning a compressed representation of the data. This approach allows the model to capture complex relationships and uncertainties in the data, improving the accuracy of disaggregation. The system may also include preprocessing steps to clean and normalize the data before feeding it into the model, as well as post-processing steps to refine the output. The invention is particularly useful in applications like energy monitoring, where disaggregated data provides insights into individual appliance usage, enabling better energy management and conservation.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
April 6, 2020
March 15, 2022
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.